An Experimental Study of Dimensionality Reduction Methods

Abstract

Dimensionality reduction (DR) lowers the dimensionality of a high-dimensional data set by reducing the number of features for each pattern. The importance of DR techniques for data analysis and visualization led to the development of a large diversity of DR methods. The lack of comprehensive comparative studies makes it difficult to choose the best DR methods for a particular task based on known strengths and weaknesses. To close the gap, this paper presents an extensive experimental study comparing 29 DR methods on 13 artificial and real-world data sets. The performance assessment of the study is based on six quantitative metrics. According to our benchmark and evaluation scheme, the methods mMDS, GPLVM, and PCA turn out to outperform their competitors, although exceptions are revealed for special cases.